[Parquet][Python] Potential regression in Parquet parallel reading

[eeroel]

Describe the enhancement requested

UPDATE: this is looking more like a bug on closer look. What happens:

When calling to_table() on a FileSystemDataset in Python using pyarrow.fs.S3FileSystem,

• Using 02de3c1, one HEAD request and two GET requests are made for each file. Also the requests are made concurrently.
• With current main, there are two HEAD requests and three GET requests for each file. Also, the first HEAD request is made from the main thread so the downloads are started sequentially. I would expect to see only one HEAD request, not sure if the three GET are expected due to some change.

Here's an example using 02de3c1, reading a FileSystemDataset using fragment_readahead = 100 and io concurrency set to 100; Y-axis represents files and X-axis is time in seconds, and each point is the relative start time of a request (HEAD or GET):

With the current main fc8c6b7 it seems that the first request for each file is made from the same thread (blue), and notably there are five requests per each file.

See comment below for reproducible example.

I'm running on Max OS 14.1.

Component(s)

Parquet

[eeroel]

As a side note, I was a bit disappointed with the total download times I achieved with maximum parallelism using 02de3c1 on my system (Macbook to S3). The requests start with nice concurrency but a few of the GET requests end up taking quite long to finish, so fragment_readahead > 4 doesn't speed up the total completion time at all. Not sure if this is related to my network, or if it's something to optimize in Arrow.

[eeroel]

I noticed that there is an extra HEAD request coming from somewhere so I think what I wrote below is unlikely to be related

Just a hunch, haven't followed the whole code path, but could it be that e.g. from here https://github.com/apache/arrow/blob/fc8c6b7dc8287c672b62c62f3a2bd724b3835063/cpp/src/arrow/dataset/file_parquet.cc#L606C72-L606C72 or EnsureCompleteMetadata the sync version of GetReader gets called, not GetReaderAsync? Introduced in 76c4a6e

EDIT: or maybe from here

arrow/cpp/src/arrow/dataset/scanner.cc

Line 1003 in fc8c6b7

FragmentsToBatches(std::move(fragment_gen), scan_options));

as I think this function gets called when invoking Python Dataset.to_table

arrow/cpp/src/arrow/dataset/scanner.cc

Line 427 in fc8c6b7

Result<EnumeratedRecordBatchGenerator> AsyncScanner::ScanBatchesUnorderedAsync(

[eeroel]

Here's a reproducible example that doesn't use FileSystemDataset but parquet.read_table:

import pyarrow._s3fs
pyarrow._s3fs.initialize_s3(pyarrow._s3fs.S3LogLevel.Trace)

from pyarrow.dataset import (
    ParquetFileFormat,
    ParquetFragmentScanOptions
)

import pyarrow
import pyarrow.fs
import pyarrow.parquet

fs = pyarrow.fs.S3FileSystem(region="us-east-2")
fs = pyarrow.fs.SubTreeFileSystem("ursa-labs-taxi-data", fs)

pyarrow.parquet.read_table("2019/01/data.parquet", filesystem=fs)

This does not only make two HEAD requests, but four of them in total (first two in getting the schema?)

cat /tmp/foo2.log | grep HeaderOut
[DEBUG] 2023-11-07 06:30:22.913 CURL [0x1e4d11ec0] (HeaderOut) HEAD /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:23.059 CURL [0x1e4d11ec0] (HeaderOut) HEAD /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:23.190 CURL [0x1e4d11ec0] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:23.613 CURL [0x1e4d11ec0] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:23.904 CURL [0x1e4d11ec0] (HeaderOut) HEAD /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:24.028 CURL [0x17017f000] (HeaderOut) HEAD /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:24.158 CURL [0x17020b000] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:24.552 CURL [0x170297000] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:26.542 CURL [0x17017f000] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1
[DEBUG] 2023-11-07 06:30:29.142 CURL [0x17020b000] (HeaderOut) GET /2019/01/data.parquet HTTP/1.1

[mapleFU]

Using 02de3c1, one HEAD request and two GET requests are made for each file. Also the requests are made concurrently.

Since this is merged recently, is that same api being used ? I was a bit confused because this part of code didn't changed frequency recently...

[eeroel]

Using 02de3c1, one HEAD request and two GET requests are made for each file. Also the requests are made concurrently.

Since this is merged recently, is that same api being used ? I was a bit confused because this part of code didn't changed frequency recently...

I'm doing a bisect at the moment, I'll update here when I'm done.

[mapleFU]

#36765

The issue related patch changes Prefetch default mode from Default to LazyDefault, but it maybe doesn't increment the req-count..

[eeroel]

#36765

The issue related patch changes Prefetch default mode from Default to LazyDefault, but it maybe doesn't increment the req-count..

This is the commit after which there is an extra HEAD request:
0793432

[eeroel]

This is probably where the extra requests come from, I wonder if these source.Open() and parquet::ParquetFileReader::OpenAsync calls are necessary?

arrow/cpp/src/arrow/dataset/file_parquet.cc

Line 507 in 0793432

ARROW_ASSIGN_OR_RAISE(auto input, source.Open());

[mapleFU]

Very nice catch! Seems that this commit ignore your change and re-create the file. this might be done before your change is checked in 😅 And after your change and rebase, this didn't get modified.

[eeroel]

Very nice catch! Seems that this commit ignore your change and re-create the file. this might be done before your change is checked in 😅 And after your change and rebase, this didn't get modified.

Yeah that was my reading as well, I can try fixing it later